Vertical stratification of forest canopy for segmentation of under-story trees within small-footprint airborne LiDAR point clouds Hamid Hamraza*, Marco A. Contrerasb, Jun Zhanga a: Department of Computer Science, b: Department of Forestry University of Kentucky, Lexington, KY 40506, USA [email protected], [email protected], [email protected] * Corresponding Author: [email protected] +1 (859) 489 1261 Abstract Airborne LiDAR point cloud representing a forest contains 3D data, from which vertical stand structure even of under-story layers can be derived. This paper presents a tree segmentation approach for multi- story stands that stratifies the point cloud to canopy layers and segments individual tree crowns within each layer using a digital surface model based tree segmentation method. The novelty of the approach is the stratification procedure that separates the point cloud to an over-story and multiple under-story tree canopy layers by analyzing vertical distributions of LiDAR points within overlapping locales. Unlike previous work that stripped stiff layers within a constrained area, the procedure stratifies the point cloud to flexible tree canopy layers over an unconstrained area with minimal over/under-segmentations of tree crowns across the layers. The procedure does not make a priori assumptions about the shape and size of the tree crowns and can, independent of the tree segmentation method, be utilized to vertically stratify tree crowns of forest canopies with a variety of stand structures. We applied the proposed approach to the University of Kentucky Robinson Forest – a natural deciduous forest with complex terrain and vegetation structure. The segmentation results showed that using the stratification procedure strongly improved detecting under-story trees (from 46% to 68%) at the cost of introducing a fair number of over-segmented under-story trees (increased from 1% to 16%), while barely affecting the segmentation quality of over- story trees. Results of vertical stratification of canopy showed that the point density of under-story canopy layers were suboptimal for performing reasonable tree segmentation, suggesting that acquiring denser LiDAR point clouds (becoming affordable due to advancements of the sensor technology and platforms) would allow more improvements in segmenting under-story trees. Keywords: remote sensing, discrete return LiDAR, multi-story stand, canopy layering, individual tree segmentation. 1 Introduction In the past two decades, airborne light detection and ranging (LiDAR) technology has extensively been used for forestry purposes due to its ability to capture data at unprecedented spatial and temporal resolutions in the shape of 3D point clouds (Ackermann 1999; Hyyppä et al. 2012; Maltamo et al. 2014; Swatantran et al. 2016; Wehr and Lohr 1999). From this data, more detailed tree level information can be retrieved to improve the accuracy of forest assessment, monitoring, and management activities (Duncanson et al. 2012; Vastaranta et al. 2011; Weinacker et al. 2004; Wulder et al. 2012). Due to the ability to penetrate vegetation canopy, LiDAR 3D point clouds also contain vertical information from which vegetation structural information even from under-story canopy layers can be retrieved (Hall et al. 2011; Lefsky et al. 2002; Maguya et al. 2014; Reutebuch et al. 2005), which is of great value for various forestry applications and ecological studies (Espírito-Santo et al. 2014; Ishii et al. 2004; Singh et al. 2015; Wing et al. 2012). Although understory trees provide limited financial value and a minor proportion of total above ground biomass, they influence canopy succession and stand development, form a heterogeneous and dynamic habitat for numerous wildlife species, and are an essential component of forest ecosystems (Antos 2009; Heurich 2008; Jules et al. 2008; Moore et al. 2007). However, to obtain individual trees attributes (e.g., location, crown width, height, DBH, volume, biomass) from different canopy layers, accurate and automated tree segmentation approaches that are able to separate tree crowns both vertically and horizontally are required (Duncanson et al. 2014; Ferraz et al. 2012; Shao and Reynolds 2006; Wang et al. 2008). Numerous methods for individual tree segmentation within LiDAR data have been developed. Earlier methods use pre-processed data in the form of digital surface models (DSMs) or canopy height models (CHMs) to segment individual trees (Chen et al. 2006; Jing et al. 2012; Koch et al. 2006; Kwak et al. 2007; Popescu and Wynne 2004; Véga and Durrieu 2011). These methods have an inherent drawback of missing under-story trees by considering only the surface data (Hamraz et al. 2016; Wang et al. 2008). More recent methods process the raw point clouds in order to utilize all horizontal and vertical information and, from the computational viewpoint, can be classified to volumetric or profiler methods. Volumetric methods directly search the 3D volume for the individual trees (Amiri et al. 2016; Ferraz et al. 2012; Lahivaara et al. 2014; Li et al. 2012; Lindberg et al. 2014; Lu et al. 2014; Rahman and Gorte 2009; Véga et al. 2014), hence are generally computationally intensive and may be prone to suboptimal solutions due to the 1 large magnitude of the search space. On the other hand, profiler methods tame the computational load through a more modular process. They typically have a module for vertical segmentation, i.e., to strip the 3D volume to multiple 2D horizontal profiles, a module for horizontal segmentation, i.e., to search the trees within the profiles, and a module to ultimately aggregating the results across the profiles (Ayrey et al. 2017). However, they generally lose information about the vertical crown geometry when processing a 2D profile. To minimize information loss due to profiling, other profiler methods have analyzed vertical distribution of LiDAR points to identify 2.5D profiles embodying more information about vertical crown geometry. Wang et al. (2008) searched trees within each profile and used a top-down routine to unify any detected crown that may be present in different profiles. They analyzed vertical distribution of all LiDAR points globally within a given area to determine the height levels for stripping profiles. However, depending on the vegetation height variability, a globally derived height level may lead to under/over-segmenting tree crowns across the profiles. Other approaches addressed this issue by identifying constrained regions including one or more trees using a preliminary segmentation routine and independently 2.5D profiling each region (Duncanson et al. 2014; Paris et al. 2016; Popescu and Zhao 2008), yet the final result is dependent on the preliminary segmentation. Although a number of methods for segmenting individual trees in multi-story stands have been proposed, they are still unable to satisfactorily detect most of the under-story trees. Typically, detection rate of dominant and co-dominant (over-story) trees is around or above 90% and detection rate of intermediate and overtopped (under-story) trees is below 50%. This inefficacy can be attributed to the reduced amount of LiDAR points penetrating below the main cohort formed by over-story trees (Kükenbrink et al. 2016; Takahashi et al. 2006), although incompetency of the current approaches to effectively use all vertical and horizontal information also plays a role. In this paper, we propose a profiler approach for segmenting crowns of all size trees in multi-story stands. The approach derives height levels locally hence stratifies the point cloud to 2.5D profiles (hereafter referred to as canopy layers), each of which is sensitive to stand height variability and includes a layer of non-overtopping tree crowns within an unconstrained area. The approach then utilizes a DSM-based method as a building block to segment individual tree crowns within each canopy layer. 2 2 Materials and Methods 2.1 Study site and LiDAR campaign The study site is the University of Kentucky’s Robinson Forest (RF, Lat. 37.4611, Long. - 83.1555) located in the rugged eastern section of the Cumberland Plateau region of southeastern Kentucky in Breathitt, Perry, and Knott counties (Figure 1). The terrain across RF is characterized by a branching drainage pattern, creating narrow ridges with sandstone and siltstone rock formations, curving valleys and benched slopes. The slopes are dissected with many intermittent streams (Carpenter and Rumsey 1976) and are moderately steep ranging from 10 to over 100% facings predominately northwest and south east, and elevations ranging from 252 to 503 meters above sea level. Vegetation is composed of a diverse contiguous mixed mesophytic forest made up of approximately 80 tree species with northern red oak (Quercus rubra), white oak (Quercus alba), yellow-poplar (Liriodendron tulipifera), American beech (Fagus grandifolia), eastern hemlock (Tsuga canadensis) and sugar maple (Acer saccharum) as over-story species. Under-story species include eastern redbud (Cercis canadensis), flowering dogwood (Cornus florida), spicebush (Lindera benzoin), pawpaw (Asimina triloba), umbrella magnolia (Magnolia tripetala), and bigleaf magnolia (Magnolia macrophylla) (Carpenter and Rumsey 1976; Overstreet 1984). Average canopy cover across RF is about 93% with small opening scattered throughout. Most areas exceed 97% canopy cover and recently harvested areas have an average cover as low as 63%. After being extensively logged in the 1920’s, RF is considered second growth